Moonshot AI Releases Kimi K3.1 Pro with Focus on Finance and Coding

Moonshot AI released Kimi K3.1 Pro on October 23. It is available the same day across the Kimi web version and app, the office assistant Kimi Work, the coding tool Kimi Code, and via API.

Officially described as their strongest open-weight model to date, full weights are scheduled for release by November 4.

K3.1 Pro was further trained from the previous generation, Kimi K3. The model architecture remains unchanged, but significant additional compute resources were dedicated to reinforcement learning—essentially having the model repeatedly perform tasks across hundreds of domains and improve based on outcome scoring.

The model has a total of 2.8 trillion parameters, activating approximately 104 billion per token (a token is the unit AI uses to process text, akin to a fragment of words). This structure is known as Mixture-of-Experts (MoE): the model contains many specialized "expert" sub-networks, with only a small subset activated for each task. This allows it to compute faster than standard models of equivalent size, though the entire model must still be loaded into memory. It supports a context window of 1 million tokens—equivalent to a medium-sized code repository or several hundred pages of financial reports—and can also process images and video.

Comparison with Claude and GPT

The following scores come from comparison tables published by Moonshot AI. Competitors include Anthropic’s Claude Opus 5.5, Claude Fable 5.1, and OpenAI’s GPT-6 Astra, all configured for maximum reasoning intensity. Some competitor scores were measured internally by Kimi; these internal tests cannot be reproduced externally.

Coding performance is largely comparable.

On Terminal-Bench 4.0, which evaluates real-world development tasks in the command line, K3.1 Pro scored 65.7%, compared to 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. On SWE-Marathon, which involves large-scale software projects, K3.1 Pro ranked first with a score of 51.7, ahead of Opus 5.5 at 46.3. However, there is a significant gap in scientific computing tasks on Terminal-Bench Science: K3.1 Pro scored 38.1, while GPT-6 Astra achieved 68.1.

Slightly behind in 3D modeling.

K3.1 Pro can directly write scripts to operate Blender, CAD software, and game engines like Unity, Unreal, and Godot. Official demos showcased the creation of complete 3D games and 3D-printable lamps. On BenchCAD, which involves reconstructing parts from rendered images using CAD code, K3.1 Pro scored 92.7%, trailing Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).

Finance is the primary focus this time.

On OfficeQA Pro v2, which involves reading real financial documents and performing multi-step calculations, K3.1 Pro ranked first with a score of 64.4, surpassing Opus 5.5 (62.2). In Kimi’s internal tests, financial modeling improved by 7.2 points over K3, and research report writing improved by 11.6 points.

The claim that "hallucinations reduced by 65%" comes from an internal fact-checking test designed specifically for difficult cases: For every 100 facts stated by K3, 23 were incorrect; for K3.1 Pro, this dropped to just over 8. In the same test, Opus 5.5 had 15 errors, and GPT-6 Astra had 5.5. The official note clarifies that this is a stress test and does not represent error rates in daily usage.

Pricing and Self-Hosting

General users can access the model via the Kimi web version and app; the specific limits for free users have not been officially disclosed. Enterprise features are disabled by default and require manual activation by administrators.

API pricing is $2 per million input tokens and$10 per million output tokens, two-thirds the cost of K3. Claude Opus 5.5 costs $4 for input and$20 for output, double the price. The Highspeed version, which is 2–3 times faster, costs $4 for input and$20 for output, matching Opus 5.5. API accounts require a minimum top-up of $1 to use.

Open weights mean the model files are publicly available for anyone to download, run on their own servers, and fine-tune without data passing through Moonshot AI. However, personal computers cannot run it: Based on the 4-bit precision used by K3, storing the 2.8 trillion parameters alone requires approximately 1.4TB of memory. License terms have not yet been published.

To support open weights, Moonshot AI stated that they removed data related to network offense and defense during training. Without removing this data, their internal vulnerability exploitation test solved 90.5% of problems; after removal, the score dropped to 81.0%. Officials claim its offensive capabilities still lag behind top-tier closed-source models that have been public for over three months.

Original Article: https://t.co/9qim2A0qJt Official Blog: https://t.co/dag6GIhBIE API Documentation: https://t.co/yRPoGN9WCo

↳ Quoting @Kimi_Moonshot (Kimi.ai): Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.

🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal. 🔹 Full-project coding: games in Unreal, Unity & Godot. 🔹 Excels at parametric 3D modeling in Blender & CAD. 🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%. 🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.

Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on